Sharing code for Kaggle Human Protein Atlas competition
Top 4% in Leaderboard
https://www.kaggle.com/c/human-protein-atlas-image-classification
Color filters for the fluorescent images
- Green (Alexa Flour 488): The protein of interest
- Blue (DAPI): Nucleus
- Red: Microtubules
- Yellow: Endoplasmic reticulum
Train sample size: 31072
Test sample size: 11702
- Nucleoplasm 12885
- Nuclear membrane 1254
- Nucleoli 3621
- Nucleoli fibrillar center 1561
- Nuclear speckles 1858
- Nuclear bodies 2513
- Endoplasmic reticulum 1008
- Golgi apparatus 2822
- Peroxisomes 53
- Endosomes 45
- Lysosomes 28
- Intermediate filaments 1093
- Actin filaments 688
- Focal adhesion sites 537
- Microtubules 1066
- Microtubule ends 21
- Cytokinetic bridge 530
- Mitotic spindle 210
- Microtubule organizing center 902
- Centrosome 1482
- Lipid droplets 172
- Plasma membrane 3777
- Cell junctions 802
- Mitochondria 2965
- Aggresome 322
- Cytosol 8228
- Cytoplasmic bodies 828
- Rods & rings 11
Resnet FastAI
python3 -m code.resnet_fastai --imagesize=256 --loss=focal --gpuid=0
python3 -m code.resnet_fastai --model=stage-2-resnet50-224-drop0.5-ep5_15 --gpuid=0
- -a, --arch : Neural network architecture
- -b, --batchsize : batch size
- -d, --encoderdepth : encoder depth of the network
- -D, --dataset : Dataset
- -e, --epochnum1 : epoch number for stage 1
- -E, --epochnum2 : epoch number for stage 2
- -f, --fold : K fold cross validation
- -i, --gpuid : GPU device id
- -l, --loss : loss function
- -m, --model : trained model to load
- -p, --dropout : dropout ratio
- -r, --learningrate : learning rate
- -s, --size : image size
- -S, --sampler : sampler
- -t, --thres : threshold
- -v, --verbose : set verbosity 0-3, 0 to turn off output (not yet implemented)
Mask RCNN
python3 -m code.mask_rcnn train --dataset=data/official --subset=train --weights=coco
python3 -m code.mask_rcnn detect --dataset=data/official --subset=test --weights=logs/nucleus20181212T0318/mask_rcnn_nucleus_0025.h5